The Application of Graph Neural Network Based on Edge Computing in English Teaching Mode Reform

Abstract

The latest developments in edge computing have paved the way for more efficient data processing especially for simple tasks and lightweight models on the edge of the network, sinking network functions from cloud to edge of the network closer to users. For the reform of English teaching mode, this is also an opportunity to integrate information technology, providing new ideas and new methods for the optimization of English teaching. It improves the efficiency of English reading teaching, stimulates the interest of English learning, enhances students’ autonomous learning ability, and creates favorable conditions for students’ learning and development. This paper designs a MEC-based GNN (GCN-GAN) user preference prediction recommendation model, which can recommend high-quality video or picture text content to the local MEC server based on user browsing history and user preferences. In the experiment, the LFU-LRU joint cache placement strategy used in this article has a cache hit rate of up to 99%. Comparing the GCN-GAN model with other traditional graph neural network models, it performs caching experiments on the Douban English book data and Douban video data sets. The GCN-GAN model has a higher score on the cache task, and the highest speculation accuracy value F1 can reach 86.7.

This work is licensed under http://creativecommons.org/licenses/by/4.0/ (the “License”). Notwithstanding the ProQuest Terms and Conditions, you may use this content in accordance with the terms of the License.

1. Introduction

In recent years, applications based on deep learning have brought great improvement to people’s lives. People are more and more interested in the extension of graph deep learning methods. With the success of many factors, a new research hotspot “Graph Neural Network (GNN)” came into being. In the era of Internet of Everything, it is a big challenge to run computational intensive deep learning algorithms on edge devices with limited resources.

Under the background of educational informationization, the introduction of advanced technology into the classroom is also an inevitable development trend. In view of the disadvantages of traditional English classroom teaching, new technology can provide new ideas and methods for the reform of English teaching mode. Traditional English reading teaching mainly has problems such as lagging teaching concepts, single teaching materials, limited teaching content, and low students’ sense of self-efficacy. With the support of information technology, English teaching mode reform is carried out to get rid of the shortcomings of traditional English teaching mode, so as to expand students’ knowledge and cultivate students’ reading ability.

This article designs a platform that can intelligently recommend learning content by combining MEC technology and graph neural network algorithms. It deepens teachers’ understanding of information-based teaching, promotes the renewal of teachers’ teaching concepts and reflections on teaching practice, and finally realizes the in-depth development of English teaching reform at the junior high school stage.

The development of various intelligent technologies has promoted the reform process of education and teaching. The innovations of this paper are as follows: First, this paper adopts the LFU-LRU joint cache placement strategy and proposes a new online joint collaborative cache and processing algorithm, which can improve the cache hit rate. The second is to introduce the graph attention mechanism into the graph convolutional neural network to complement the functions to improve the performance of the graph neural network model. Thirdly, based on edge computing and machine learning technology, an adaptive recommendation system for English teaching oriented to users and content is constructed.

2. Related Work

In order to reduce the transmission delay of network services, edge computing has attracted more and more attention from the industry and academia, and due to the continuous development of deep recognition technology, running deep learning algorithms on edge devices with limited resources has become a research hotspot. Zhang Y. et al. proposed a computing resource allocation scheme for IoV mobile edge computing scenarios based on deep reinforcement learning network. They determined the task resource allocation model in the corresponding edge computing scenario, with the minimum total computing cost as the objective function and established the mathematical model of task offloading and resource allocation [1]. The resource allocation model they established can indeed effectively cope with the problem of slow data return, but the research does not make empirical application of the model’s use effect in practical applications; it is only a theoretical explanation. Zhang J. et al. proposed a new data layout method for edge-oriented computing vector processor with specific neural network model and applied it to feature mapping. They proposed a method of parallelizing matrix convolution calculation in three-dimensional space to improve access efficiency [2]. The data layout method proposed by Zhang J.’s research can greatly improve the efficiency of edge computing for data access, but its use of neural network algorithms needs to be further optimized. Based on the graph convolutional neural network, Ahmad et al. proposes a graph sparsity technique that uses effective edge resistance to better model global context information and eliminate redundant nodes and edges in the graph. In addition, they combined self-attention graph pools to preserve local attributes [3]. Xu proposes an edge computing based on a deep reasoning framework, which has the privacy of local differences in mobile data analysis. The deep learning model is used to minimize data and adaptively inject noise to confuse the learned features, thereby forming a new protective layer to resist sensitive inference [4]. His research is mainly for edge computing, and there is certain research progress, but the feasibility of the scheme used has yet to be verified. Liu and He developed and designed a system software based on cognitive computing in the embedded ARM server system and also built the related system database. This system is mainly used in the reform of teaching technology, which can improve the intelligence of teaching methods and achieve low error rate in the transmission of teaching resources [5]. Its research is a great progress for the application of artificial intelligence in the reform of teaching mode, and it also provides new ideas for the research of this article. Wei et al. studied the computational offloading problem of mobile users in mobile edge computing wireless cellular networks and used a model-free reinforcement learning (RL) framework to describe and solve the computational offloading problem. Each mobile user interacts with the environment and chooses local computing or edge computing according to its status [6]. The research is mainly aimed at the mobile user computing offloading problem of mobile edge computing and provides a more adaptable solution. However, the research process is too complicated, and the actual application effect may not be good.

3. The Application Research Method of Graph Neural Network Based on Edge Computing in the Reform of English Teaching Mode

3.1. Edge Computing Technology

Edge computing is a new technology for the Internet of Things, which lies between physical entities and industrial connections. Its model diagram is shown in Figure 1. It can be seen that it is an open platform for users and content, which starts the response at the edge of the cloud platform, can meet the service requirements of various organizations and industries in real-time business, intelligent application, and data security protection, and has faster response speed and convenient access than the cloud platform [7, 8]. In essence, edge computing is an application mode that integrates the core technologies of network, computing, storage, and application to provide the nearest service on the side close to things or data sources.

[figure omitted; refer to PDF]

This model combines the characteristics of the spatial structure of the spectrum and graph topology in GCN and approximates the Laplacian matrix: $$\begin{array}{c}\text{(9)}& Y^{l+1}=\delta {\stackrel{\wedge }{E}}^{-\frac{1}{2}}\widehat{Q}{\stackrel{\wedge }{E}}^{-\frac{1}{2}}\ast Y^l{W}^l+a^l,\widehat{Q}=Q+I,\\ {\widehat{E}}_{ii}=\sum _j{\widehat{E}}_{ij},\end{array}$$

Among them, $\delta$ represents the activation function, $E$ represents the matrix of the node, $Q$ node represents the adjacency matrix, $I$ is the identity matrix, l is the number of convolutional layers, and $W^l$ and $a^l$ are the training parameters and influence factors.

3.2.2. Graph Attention Mechanism (GAN)

(1) Graph Attention Network (GAT). This is a space-based graph convolutional network. The attention mechanism uses the attention mechanism to determine the weight of the node neighborhood when aggregating feature information. Its operation is defined as $$\begin{array}{}\text{(10)}& H_i^t=\sigma \sum _{j\in M_t}\alpha H_i^{t-1},H_j^{t-1}{W}^{t-1}{H}_j^{t-1},\end{array}$$where $\alpha$ represents an attention function that can adaptively control the contribution of adjacent nodes $j$ to $i$. GAT can also use a multiattention mechanism to deal with the weights of different subspaces: $$\begin{array}{}\text{(11)}& {H_i^t=}_{f=1}^F\sigma \sum _{j\in M_t}{\alpha }_f{H}_i^{t-1},H_j^{t-1}{W}^{t-1}{H}_j^{t-1}.\end{array}$$

(2) Graph Attention Model (GAM). In order to improve the system’s ability to classify and identify graph data, the model provides a recurrent neural network model. The main purpose is to achieve adaptive access to the important node sequence for processing graph information. The model is defined as $$\begin{array}{}\text{(12)}& H_t=g_H{g}_k{R}_{t-1},S_{t-1},f;{\vartheta }_k,H_{t-1};{\vartheta }_h,\end{array}$$where $g_H$ is an LSTM network and $g_k$ is a step network. Its function is to give priority access to neighbors with high priority of the current node and aggregate their information.

The advantage of GAT and GAAN is that they can adaptively learn the importance weights of neighbor nodes. However, the computational cost and memory consumption increase rapidly with the calculation of the attention weight between each pair of neighbors.

(3) GCN-GAN Joint Graph Neural Network Algorithm. The performance of GCN algorithm is excellent among many algorithms. However, in order to improve the information extraction effect of the structure in the downsampling mechanism, this research introduces the graph self-attention mechanism into the pooling mechanism of the graph convolution model, which can give higher weight to the information data that reflects the category. The specific model is shown in Figure 5. The model uses top-k interception in the pooling operation and uses the mask mechanism to filter nodes other than top-k. The advantage of this model is that it fully considers the topological structure of the node in the network and the inherent characteristic information of the node during the pooling process. The calculation principle of the structure output is $$\begin{array}{c}\text{(13)}& QI=\sigma {\stackrel{\wedge }{D}}^{-\frac{1}{2}}\widehat{A}{\stackrel{\wedge }{D}}^{-\frac{1}{2}}\ast H{W}_{\text{ztt}}+b_{\text{att}},\text{idx}=\text{Top}-\mathrm{rank}Q,\text{kN},\\ Q_{\text{mask}}=Q_{\text{idx}},\\ X_{\text{out}}=X_{\text{idx}},:⨀Q_{\text{mask}},\\ A_{\text{out}}=A_{\text{idx},\text{idx}},\end{array}$$where $H$ represents a node vector, $W_{\text{ztt}}$ and $b_{\text{att}}$ are the model parameter variables, $Q_{\text{mask}}$ is a Boolean matrix, k is the proportion of pooled nodes, $X_{\text{out}}$ is a new node feature matrix, and $A_{\text{out}}$ is an adjacency matrix.

[figure omitted; refer to PDF]

The target loss function selected in this model is a binary cross entropy function, which is optimized by the loss function of the variational automatic encoder module, namely $$\begin{array}{}\text{(14)}& \text{loss}=\text{BCE}y,\widehat{y}+\beta \times 0.5\times -\log {\sigma }^2+{\sigma }^2+{\mu }^2-1.\end{array}$$

3.3. English Teaching Application of Graph Neural Network Based on Mobile Edge Computing

3.3.1. Disadvantages of Traditional English Teaching Mode

Teaching mode is the system structure of education and teaching methods constructed to achieve certain teaching objectives. It is based on certain education and teaching theory and practical experience. It is an intermediary that transforms relevant teaching theories into specific teaching activities and operating procedures and is the result of combining relevant teaching theories and practical frameworks with specific teaching situations.

This research takes the current situation of Chinese junior middle school English teaching mode as an example to illustrate. For a long time, China has mostly adopted the traditional “top-down” reading teaching model for English teaching. This model often adopts the method of word-for-word translation and comprehension of the article. It only sees the trees but not the forest. There is an obvious phenomenon of “language rather than culture.” It is difficult for students to simply comprehend words to stimulate their interest in learning and to comprehend the true meaning of language, which has many disadvantages.

The gradual development of interactive English teaching mode provides a favorable opportunity for the reform of English teaching mode. Multimedia, computer network technology, and some artificial intelligence technologies have played an important role in education and teaching activities. It is important and necessary to carry out informatization teaching and then to promote the exploration of English teaching reform, which can further cultivate students’ autonomous reading ability.

3.3.2. English Teaching Content Recommendation Based on Edge Computing-Graph Neural Network

The purpose of this article is to build a teaching platform around the difficulties of English teaching in junior high schools, supported by edge computing technology and graph neural network algorithms. According to the browsing records of pictures or videos of English knowledge by students and teachers, it can adaptively recommend English teaching content that students or teachers are interested in. The platform architecture is shown in Figure 6. Edge computing can quickly obtain user browsing data and cache the recommended content for the next time period locally. The graph neural network model can learn the characteristics of user browsing content and then can classify the content and can more accurately recommend relevant learning content to users [20, 21].

[figure omitted; refer to PDF]

Figure 8 shows the performance comparison of the three caching mechanisms on ADL and HR under the MEC cache space change. The skew coefficient is 0.68, the range of MEC buffer space is set at 10-100, and the number of service contents in the network system is 1200. Under the three caching mechanisms, the results of the impact of cache space size changes on cache performance are given in Figure 8, namely, the trends of ADL and HR. Experimental results show that the average transmission delay of these cache mechanisms decreases with the increase of cache space, and the cache hit rate increases with the increase of cache space. Specifically, when the cache space is 50, the cache hit rate under this mechanism is 11.8% and 5.6% higher than that of LRU and LFU, respectively.

[figure omitted; refer to PDF]

The cache space size of each experimental MEC server is 50, and the setting range of the number of service contents is 500 to 5000. The comparison results of the caching performance of the three caching mechanisms under the change of service contents are shown in Figure 9. Among the three caching strategies, the LFU-LRU caching mechanism proposed in this article is better than the other two. In summary, the LFU-LRU caching mechanism proposed in this paper can effectively improve the cache hit rate and reduce transmission overhead and content transmission delay.

[figure omitted; refer to PDF]

4.2. MEC—GNN English Teaching Content Recommendation Platform Performance Experiment

4.2.1. Introduction to the Experimental Data Set

The experimental data set is divided into three categories. The first category is the miniImagNet data set, which is used to test the content classification performance of the platform. This data set is an excerpt from the ImageNet data set and is a small classification data set. The second and third types of data are Douban English book data and Douban video data, respectively, used to test the prediction accuracy of user preference transfer.

4.2.2. Experimental Results

This section of the experiment is mainly to test the performance of the English teaching content recommendation platform based on the MEC-based GNN (GCN-GAN) neural network proposed in this paper. It will be compared with some other deep learning methods and the situation of evaluating the target domain vector with three cache indicators of ROC-AUC, PR-AUC, and F1. The experimental results on the miniImagNet data set are shown in Table 3. It can be intuitively seen from Table 3 that the graph convolution-self-attention mechanism model based on edge computing proposed in this paper is the best for the classification of the data set. ROC-AUC, PR-AUC, and F1 values are higher than the corresponding experimental values of GAT, GCN, and DeepWalk.

Table 3

The performance of different models on the miniImagNet data set.

	ROU—AUC	PR—AUC	F1
GAT	78.16	79.02	80.28
GCN	81.65	82.87	83.01
DeepWalk	76.44	79.25	78.05
GCN—GAN	83.52	84.71	85.59

The caching experiment results of several prediction models on the Douban English book data and Douban video data sets are shown in the left and right images of Figure 10, respectively. It can be seen from the experimental results that in this task, the user preference prediction and recommendation effect of English learning content based on learning transfer in this paper are the best. The other types of graph neural network models do not directly process graph structures in nature. Therefore, the performance score on the cache task is not too high. The GCN-GAN model has a higher score on the cache task, and the inference process is more efficient. The highest inference accuracy value F1 can reach 86.7.

[figure omitted; refer to PDF]

5. Discussion

In this experiment, a cooperative caching mechanism for edge computing based on machine learning graph neural network is proposed. And it uses the method of transfer learning to predict the user’s preference for push content. The experimental results show that the LFU-LRU caching strategy proposed in this research can effectively improve the cache hit rate and reduce the content transmission delay. The collaborative caching mechanism of graph neural network based on mobile edge computing in this paper is for the reform of English teaching mode, so the content of cache prediction is the English learning-related content browsed by teachers or students in the wireless network domain. In the prediction accuracy simulation experiment, the article compares the user preference prediction mechanism proposed in this paper with the other deep learning graph neural network models. Using ROC-AUC, PR-AUC, and F1, three cache indicators to evaluate the evaluation of the target domain vector, the results show that the optimized GCN-GAN model in this paper has a higher score on the cache task, and the inference process has a higher efficiency.

6. Conclusions

This article focuses on English teaching as the research object, using 5G mobile edge computing technology (MEC) and the graph neural network algorithm in the deep learning algorithm as modern technical support, constructing a wireless network cooperative caching mechanism. It can cache and recommend relevant meaningful learning content to users relatively accurately. This article proposes to integrate information technology into the whole process of English teaching, effectively improve students’ comprehensive language ability, promote students’ deep learning, and further verify the effect of junior high school English reading teaching through teaching practice. It is expected to provide guidance and reference for relevant practical research and promote the continuous improvement of new methods, new experiences, and new models of junior high school English teaching under the background of educational informationization.

References

[1] Y. Zhang, M. Zhang, C. Fan, F. Li, B. Li, "Computing resource allocation scheme of IOV using deep reinforcement learning in edge computing environment," EURASIP Journal on Advances in Signal Processing, vol. 2021,DOI: 10.1186/s13634-021-00750-6, 2021.

[2] J. Zhang, Y. Guo, X. Hu, R. Li, "Design and implementation of deep neural network for edge computing," IEICE Transactions on Information and Systems, vol. E101.D no. 8, pp. 1982-1996, DOI: 10.1587/transinf.2018EDP7044, 2018.

[3] T. Ahmad, L. Jin, L. Lin, G. Z. Tang, "Skeleton-based action recognition using sparse spatio-temporal GCN with edge effective resistance," Neurocomputing, vol. 423 no. 4, pp. 389-398, DOI: 10.1016/j.neucom.2020.10.096, 2021.

[4] C. Xu, J. Ren, L. She, Y. Zhang, Z. Qin, K. Ren, "EdgeSanitizer: locally differentially private deep inference at the edge for mobile data analytics," IEEE Internet of Things Journal, vol. 6 no. 3, pp. 5140-5151, DOI: 10.1109/JIOT.2019.2897005, 2019.

[5] P. H. Liu, S. B. He, "Analysis of teaching reform mode based on cognitive computing system - an example of dragon boat teaching," Cognitive Systems Research, vol. 52 no. DEC., pp. 978-984, DOI: 10.1016/j.cogsys.2018.09.013, 2018.

[6] Y. Wei, Z. Wang, D. Guo, F. Richard Yu, "Deep Q-learning based computation offloading strategy for mobile edge computing," Computers, Materials and Continua, vol. 59 no. 1, pp. 89-104, DOI: 10.32604/cmc.2019.04836, 2019.

[7] Y. Zhang, Z. Sun, P. Jiang, Y. Huang, J. Peng, "Hybrid image retargeting using optimized seam carving and scaling," Multimedia Tools and Applications, vol. 76 no. 6, pp. 8067-8085, DOI: 10.1007/s11042-016-3318-1, 2017.

[8] J. Wang, J. Hu, G. Min, W. Zhan, Q. Ni, N. Georgalas, "Computation offloading in multi-access edge computing using a deep sequential model based on reinforcement learning," IEEE Communications Magazine, vol. 57 no. 5, pp. 64-69, DOI: 10.1109/MCOM.2019.1800971, 2019.

[9] B. Li, P. Chen, H. Liu, W. Guo, X. Cao, J. du, C. Zhao, J. Zhang, "Random sketch learning for deep neural networks in edge computing," Nature Computational Science, vol. 1 no. 3, pp. 221-228, DOI: 10.1038/s43588-021-00039-6, 2021.

[10] T. Zanotti, F. M. Puglisi, P. Pavan, "Reliability and performance analysis of logic-in-memory based binarized neural networks," IEEE Transactions on Device and Materials Reliability, vol. 21 no. 2, pp. 183-191, DOI: 10.1109/TDMR.2021.3075200, 2021.

[11] W. Song, S. Wang, B. Yang, Y. Lu, X. Zhao, X. Liu, "Learning node and edge embeddings for signed networks," Neurocomputing, vol. 319 no. NOV.30, pp. 42-54, DOI: 10.1016/j.neucom.2018.08.072, 2018.

[12] X. Zhang, J. Lu, D. Li, "Confidential information protection method of commercial information physical system based on edge computing," Neural Computing and Applications, vol. 33 no. 3, pp. 897-907, DOI: 10.1007/s00521-020-05272-0, 2021.

[13] Y. You, Y. Zhao, "A human pose estimation algorithm based on the integration of improved convolutional neural networks and multi-level graph structure constrained model," Personal and Ubiquitous Computing, vol. 23 no. 3-4, pp. 607-616, DOI: 10.1007/s00779-019-01255-8, 2019.

[14] M. Inoue, S. Mizoguchi, A. Fujimaki, "Study on the force and potential for the vortex motion around the concave film edge by the numerical calculation," IEEE Transactions on Applied Superconductivity, vol. 27 no. 4,DOI: 10.1109/TASC.2016.2636249, 2017.

[15] P. Nikolaou, P. Menounou, "Analytical and numerical methods for efficient calculation of edge diffraction by an arbitrary incident signal," The Journal of the Acoustical Society of America, vol. 146 no. 5, pp. 3577-3589, DOI: 10.1121/1.5134065, 2019.

[16] J. Y. Lee, B. S. Kim, "Study on approximate curvature recovery of surface mesh for three-dimensional flow calculation," Journal of Computational Fluids Engineering, vol. 22 no. 3, pp. 54-59, DOI: 10.6112/kscfe.2017.22.3.054, 2017.

[17] L. Espig, A. Frieze, M. Krivelevich, "Elegantly colored paths and cycles in edge colored random graphs," SIAM Journal on Discrete Mathematics, vol. 32 no. 3, pp. 1585-1618, DOI: 10.1137/15M1047106, 2018.

[18] J. Li, S. Yang, Z. Qi, G. Zhao, B. Yin, F. Mo, "A calculation model for water breakthrough time of gas wells in gas reservoirs with edge water considering interlayer heterogeneity: a case study of the Lower Triassic Feixianguan gas reservoirs in the Puguang gas field," Natural Gas Industry B, vol. 7 no. 6, pp. 631-638, DOI: 10.1016/j.ngib.2020.04.005, 2020.

[19] S. Yin, S. Cui, "5g edge computing and hospital built-in cache technology application in the evaluation of elderly post-anesthesia care," Arabian Journal of Geosciences, vol. 14 no. 17, 2021.

[20] X. Zhang, N. Wang, Y. Huang, "Calculation and analysis on crest thickness of edge tooth of height-modified TA worm," Mechanism and Machine Science,DOI: 10.1007/978-981-10-2875-5, 2017.

[21] H. C. Ke, H. Wang, H. W. Zhao, W. J. Sun, "Deep reinforcement learning-based computation offloading and resource allocation in security-aware mobile edge computing," Wireless Networks, vol. 27 no. 5, pp. 3357-3373, DOI: 10.1007/s11276-021-02643-w, 2021.